def assign_all(self,
anchors,
gt_boxes,
anchors_mask=None,
gt_classes=None,
matched_thresholds=None,
unmatched_thresholds=None,
importance=None):
if anchors_mask is not None:
prune_anchor_fn = lambda _: np.where(anchors_mask)[0]
else:
prune_anchor_fn = None
def similarity_fn(anchors, gt_boxes):
anchors_rbv = anchors[:, [0, 1, 3, 4, 6]]
gt_boxes_rbv = gt_boxes[:, [0, 1, 3, 4, 6]]
return self._sim_calcs[0].compare(anchors_rbv, gt_boxes_rbv)
def box_encoding_fn(boxes, anchors):
return self._box_coder.encode(boxes, anchors)
return create_target_np(anchors,
gt_boxes,
similarity_fn,
box_encoding_fn,
prune_anchor_fn=prune_anchor_fn,
gt_classes=gt_classes,
matched_threshold=matched_thresholds,
unmatched_threshold=unmatched_thresholds,
positive_fraction=self._positive_fraction,
rpn_batch_size=self._sample_size,
norm_by_num_examples=False,
box_code_size=self.box_coder.code_size,
gt_importance=importance)
def assign(self,
anchors,
gt_boxes,
anchors_mask=None,
gt_classes=None,
matched_thresholds=None,
unmatched_thresholds=None):
if anchors_mask is not None:
prune_anchor_fn = lambda _: np.where(anchors_mask)[
0] # 返回非0元素的坐标,这里的lambda很多余,实际上不需要任何参数
else:
prune_anchor_fn = None
def similarity_fn(anchors, gt_boxes): # 根据鸟瞰图计算IOU,嵌套函数第一时间不执行
anchors_rbv = anchors[:, [0, 1, 3, 4, 6]]
gt_boxes_rbv = gt_boxes[:, [0, 1, 3, 4, 6]]
return self._region_similarity_calculator.compare(
anchors_rbv, gt_boxes_rbv)
def box_encoding_fn(boxes, anchors): # 锚框编码,嵌套函数可以第一时间不传参数
return self._box_coder.encode(boxes, anchors)
return create_target_np( # 创建训练目标的核心函数
anchors, # (248*216*2,7)
gt_boxes, # (gt_num, 7)
similarity_fn, # func
box_encoding_fn, # func
prune_anchor_fn=prune_anchor_fn, # func
gt_classes=gt_classes, # (gt_num,)从1开始标志不同真值的类别
matched_threshold=matched_thresholds, # 0.6,(248*216*2,)
unmatched_threshold=unmatched_thresholds, # 0.45,(248*216*2,)
positive_fraction=self._positive_fraction, # None
rpn_batch_size=self._sample_size, # 512
norm_by_num_examples=False,
box_code_size=self.box_coder.code_size) # 7
def assign(self,
anchors,
gt_boxes,
anchors_mask=None,
gt_classes=None,
matched_thresholds=None,
unmatched_thresholds=None):
if anchors_mask is not None:
prune_anchor_fn = lambda _: np.where(anchors_mask)[0]
else:
prune_anchor_fn = None
def similarity_fn(anchors, gt_boxes):
anchors_rbv = anchors[:, [0, 1, 3, 4, 6]] ## anchors的第7位是？
gt_boxes_rbv = gt_boxes[:, [0, 1, 3, 4, 6]]
return self._region_similarity_calculator.compare(
anchors_rbv, gt_boxes_rbv)## 计算rbv锚框的相似度
def box_encoding_fn(boxes, anchors):
return self._box_coder.encode(boxes, anchors)## box和锚框需要编码成？
return create_target_np(
anchors,
gt_boxes[gt_classes == class_name],
similarity_fn,
box_encoding_fn,
prune_anchor_fn=prune_anchor_fn,
gt_classes=gt_classes[gt_classes == class_name],
matched_threshold=matched_thresholds,
unmatched_threshold=unmatched_thresholds,
positive_fraction=self._positive_fraction,
rpn_batch_size=self._sample_size,
norm_by_num_examples=False,
box_code_size=self.box_coder.code_size)
def assign_v2(self,
anchors_dict,
gt_boxes,
anchors_mask=None,
gt_classes=None,
gt_names=None):
if anchors_mask is not None:
prune_anchor_fn = lambda _: np.where(anchors_mask)[0]
else:
prune_anchor_fn = None
def similarity_fn(anchors, gt_boxes):
anchors_rbv = anchors[:, [0, 1, 3, 4, 6]]
gt_boxes_rbv = gt_boxes[:, [0, 1, 3, 4, 6]]
return self._region_similarity_calculator.compare(
anchors_rbv, gt_boxes_rbv)
def box_encoding_fn(boxes, anchors):
return self._box_coder.encode(boxes, anchors)
targets_list = []
for class_name, anchor_dict in anchors_dict.items():
mask = np.array([c == class_name for c in gt_names], dtype=np.bool_)
targets = create_target_np(
anchor_dict["anchors"].reshape(-1, self.box_coder.code_size),
gt_boxes[mask],
similarity_fn,
box_encoding_fn,
prune_anchor_fn=prune_anchor_fn,
gt_classes=gt_classes[mask],
matched_threshold=anchor_dict["matched_thresholds"],
unmatched_threshold=anchor_dict["unmatched_thresholds"],
positive_fraction=self._positive_fraction,
rpn_batch_size=self._sample_size,
norm_by_num_examples=False,
box_code_size=self.box_coder.code_size)
targets_list.append(targets)
feature_map_size = anchor_dict["anchors"].shape[:3]
targets_dict = {
"labels": [t["labels"] for t in targets_list],
"bbox_targets": [t["bbox_targets"] for t in targets_list],
"bbox_outside_weights": [t["bbox_outside_weights"] for t in targets_list],
}
targets_dict["bbox_targets"] = np.concatenate([v.reshape(*feature_map_size, -1, self.box_coder.code_size) for v in targets_dict["bbox_targets"]], axis=-2)
targets_dict["bbox_targets"] = targets_dict["bbox_targets"].reshape(-1, self.box_coder.code_size)
targets_dict["labels"] = np.concatenate([v.reshape(*feature_map_size, -1) for v in targets_dict["labels"]], axis=-1)
targets_dict["bbox_outside_weights"] = np.concatenate([v.reshape(*feature_map_size, -1) for v in targets_dict["bbox_outside_weights"]], axis=-1)
targets_dict["labels"] = targets_dict["labels"].reshape(-1)
targets_dict["bbox_outside_weights"] = targets_dict["bbox_outside_weights"].reshape(-1)
return targets_dict
def assign_per_class(self,
anchors_dict,
gt_boxes,
anchors_mask=None,
gt_classes=None,
gt_names=None,
importance=None):
"""this function assign target individally for each class.
recommend for multi-class network.
"""
def box_encoding_fn(boxes, anchors):
return self._box_coder.encode(boxes, anchors)
targets_list = []
anchor_loc_idx = 0
anchor_gene_idx = 0
for class_name, anchor_dict in anchors_dict.items():
def similarity_fn(anchors, gt_boxes):
anchors_rbv = anchors[:, [0, 1, 3, 4, 6]]
gt_boxes_rbv = gt_boxes[:, [0, 1, 3, 4, 6]]
return self._sim_calcs[anchor_gene_idx].compare(
anchors_rbv, gt_boxes_rbv)
mask = np.array([c == class_name for c in gt_names],
dtype=np.bool_)
feature_map_size = anchor_dict["anchors"].shape[:3]
num_loc = anchor_dict["anchors"].shape[-2]
if anchors_mask is not None:
anchors_mask = anchors_mask.reshape(-1)
a_range = self.anchors_range(class_name)
anchors_mask_class = anchors_mask[
a_range[0]:a_range[1]].reshape(-1)
prune_anchor_fn = lambda _: np.where(anchors_mask_class)[0]
else:
prune_anchor_fn = None
# print(f"num of {class_name}:", np.sum(mask))
targets = create_target_np(
anchor_dict["anchors"].reshape(-1, self.box_ndim),
gt_boxes[mask],
similarity_fn,
box_encoding_fn,
prune_anchor_fn=prune_anchor_fn,
gt_classes=gt_classes[mask],
matched_threshold=anchor_dict["matched_thresholds"],
unmatched_threshold=anchor_dict["unmatched_thresholds"],
positive_fraction=self._positive_fraction,
rpn_batch_size=self._sample_size,
norm_by_num_examples=False,
box_code_size=self.box_coder.code_size,
gt_importance=importance)
# print(f"num of positive:", np.sum(targets["labels"] == self.classes.index(class_name) + 1))
anchor_loc_idx += num_loc
targets_list.append(targets)
anchor_gene_idx += 1
targets_dict = {
"labels": [t["labels"] for t in targets_list],
"bbox_targets": [t["bbox_targets"] for t in targets_list],
"importance": [t["importance"] for t in targets_list],
}
targets_dict["bbox_targets"] = np.concatenate([
v.reshape(-1, self.box_coder.code_size)
for v in targets_dict["bbox_targets"]
],
axis=0)
targets_dict["bbox_targets"] = targets_dict["bbox_targets"].reshape(
-1, self.box_coder.code_size)
targets_dict["labels"] = np.concatenate(
[v.reshape(-1) for v in targets_dict["labels"]], axis=0)
targets_dict["importance"] = np.concatenate(
[v.reshape(-1) for v in targets_dict["importance"]], axis=0)
targets_dict["labels"] = targets_dict["labels"].reshape(-1)
targets_dict["importance"] = targets_dict["importance"].reshape(-1)
return targets_dict
